Effective Multi-View for Human Activity Recognition on Skeletal Model

Abstract

The recognition of 3D human pose from 2D joint location is fundamental to numerous vision issues in analysis of video sequences. Various methods using with skeletal model have been described in past decades, but there is required a powerful system with stable and reliable manner in activity recognition because video sequences can contain different people that may be any position or scale and complex spatial interference. With the development of deep learning, skeleton-based human representation is more reliable to motion speed and appearance of human body scale. Skeleton data contains compact information of the major body joints and that support multi-view to human activity recognition. To satisfy our aim, the proposed system is developed by using OpenPose detector that achieve effective results for 2D pose and Deep Learning based approach. Our goal is to extract valuable information between human joints and to recognize correct activity from human representation in video sequences.